The Effects of Natural Disasters on Households Preferences and Behaviours: Evidence from Cambodian Rice Farmers After the 2011 Mega Flood

Transcription

1 Chapter 4 The Effects of Natural Disasters on Households Preferences and Behaviours: Evidence from Cambodian Rice Farmers After the 2011 Mega Flood Sommarat Chantarat Chulalongkorn Univerisity Kimlong Chheng Crawford School of Public Policy, Australian National University Kim Minea Agricultural and Rural Development Sothea Oum Economic Research Institute for the ASEAN and East Asia Krislert Samphantharak School of International Relations and Pacific Studies, University of California Vathana Sann Agricultural and Rural Development March 2015 This chapter should be cited as Chantarat, S., K. Chheng, K. Minea, S. Oum, K. Samphantharak and V. Sann (2015), The Effects of Natural Disasters on Households Preferences and Behaviours: Evidence from Cambodian Rice Farmers After the 2011 Mega Flood, in Sawada, Y. and S. Oum (eds.), Disaster Risks, Social Preferences, and Policy Effects: Field Experiments in Selected ASEAN and East Asian Countries, ERIA Research Project Report FY2013, No.34.Jakarta: ERIA, pp

2 CHAPTER 4 The Effects of Natural Disasters on Households Preferences and Behaviours: Evidence from Cambodian Rice Farmers After the 2011 Mega Flood Sommarat Chantarat * Chulalongkorn Univerisity Kimlong Chheng Crawford School of Public Policy, Australian National University Kim Minea Agricultural and Rural Development Sothea Oum Economic Research Institute for the ASEAN and East Asia Krislert Samphantharak School of International Relations and Pacific Studies, University of California Vathana Sann Agricultural and Rural Development This paper studies the impacts of the 2011 mega flood on preferences, subjective expectations, and behavioural choices among Cambodian rice-farming households. We found flood victims to have larger risk aversion and altruism, and lower impatience and trust of friends and local governments. The disaster also induced flooded households to adjust upward their subjective expectations of future floods and of natural resources as a safety net. Mediating (partially if not all) through these changes in preferences and expectations, the 2011 flood also affected households behavioural choices, some of which could further determine long-term economic development and resilience to future floods. We found flooded households to have lower productive investment, to substitute away social insurance with by increasing * The corresponding authors can be contacted at sommarat.chantarat@anu.edu.au and krislert@ucsd.edu. We thank our excellent survey team from the Cambodian Royal University of Agriculture and Cambodian Mekong University. Kakda Kuy and Vanna Meas from the Council for Agricultural and Rural Development provided excellent fieldwork supervision. Financial support from the Economic Research Institute for ASEAN and East Asia is gratefully acknowledged. 85

3 self-insurance and demand for market-based instruments, and more importantly, to increase the use of natural resources as insurance. These findings shed light on the design of incentive-compatible safety nets and development interventions. 1. Introduction Natural disasters often create adverse impacts on the livelihoods of people, especially those living in developing economies where access to safety nets is limited. Disasters not only destroy physical, human, and social capital of households, catastrophic disasters can lead to a change in risk, time, and social preferences. 1 In addition, largely unexpected and rare disasters as well as the success or failure of safety net institutions in coping with disasters may lead to a revision of subjective expectations of future events. Such impacts could induce changes in behavioural choices that could in turn affect long-term economic development and resilience to future floods. Understanding these consequences also has crucial policy implications for the design of incentivecompatible safety nets and development programmes for agricultural households in rural economies. This paper aims to make a contribution to the growing literature on the impacts of catastrophic events (natural disasters or civil conflicts) on household preferences and behaviours by studying the consequences of the 2011 mega flood in Cambodia the country s biggest flood in recent history on preferences, subjective expectations, and behavioural choices of affected Cambodian rice-farming households. We use the 2011 mega flood as a natural experiment and utilise discontinuity generated by this flood to create variations in flood exposure across sampled villages and households. Field surveys and experiments were used to elicit key preferences, expectations and behavioural choices. The Cambodian 2011 mega flood was a unique natural disaster event. Although flood is the most common natural disaster in Southeast Asia, most floods occur 1 Recent studies provide empirical evidence that natural disasters can cause changes in risk, time, and social preferences. For risk preference, see Eckel, et al. (2009); Cameron and Shah (2012); Cassar, et al. (2011); and Page, et al. (2012). For time preference, see Callen (2011). For social preference, see Castillo and Carter (2011); and Cassar, et al. (2011). 86

4 in Indonesia, the Philippines, and Thailand, while Cambodia has experienced relatively less frequent floods only 15 occurrences during However, unlike other countries in Southeast Asian, the death toll per flood event in Cambodia is the highest in the region, averaging nearly 90 casualties, i.e., a death toll nearly twice as high as in Indonesia and Thailand on a perevent basis. 2 The 2011 flood was particularly important since it was the largest and deadliest in recent decades, with a death toll nearly three times as high as the historical average. Heavy rain and overflow of the Mekong River and the Tonle Sap from the second week of August 2011 affected 18 out of 24 provinces in Cambodia. Impacts were especially severe among the rice farming communities, who tend to be poorer and more flood-prone. The flood caused 250 deaths, and more than 1.7 million people affected. More than 400,000 hectares (ha) of rice crops were affected, of which almost 230,000 ha (9.3 percent of the cultivated area) were severely damaged or destroyed. Moreover, 1,675 livestock were lost, and more than 70,000 drinking water wells were contaminated. It was estimated that the floods caused USD 625 million worth of losses and damage, with infrastructure damage estimated at USD 376 million. The damage included roads (national, provincial, and rural), irrigation facilities, water supply and sanitation facilities, schools, and health centres. The flooding posed a serious challenge to development and the livelihoods of people, particularly the poor and socially disadvantaged such as women and children. Given its rarity and severity, the 2011 mega flood serves as an ideal natural experiment for a study of how a disaster affects households preferences and behaviours. This study focuses particularly on the effects of the flood on ricefarming households because most of the areas directly affected by the flood in Cambodia were farmland, especially for rice cultivation, and these farms were operated by relatively poor households whose access to risk management and risk coping mechanisms was relatively limited. The mega flood therefore had substantial impacts on the livelihoods of many farming households and thus understanding these impacts would provide important insights for policymaking regarding safety nets of poor and vulnerable households. 2 These statistics are based on the Emergency Events Database (EM-DAT), one of the most comprehensive databases on disasters, maintained by the Centre for Research on the Epidemiology of Disasters (CRED) at the University of Louvain (Belgium). See Samphantharak (2014) for more details. 87

5 We found that the mega flood seemed to have made the affected Cambodian rice-farming households become more risk averse, and this increase in risk aversion appears greatest among poorer households. The mega flood also reduced impatience and increased altruistic behaviour among the affected households. Surprisingly, the 2011 flood, caused a significant reduction in trust of neighbours and local governments. Flood victims revised upward their subjective expectations of future severe floods and of the benefits of natural resources as a safety net. Mediating (partially if not all) through these changes in preferences and expectations, the 2011 flood also affected households behavioural choices. We found the flooded households to have lower productive investment, to substitute away social insurance with an increase in self-insurance and demand for market-based instruments, and more importantly, to increase the use of natural resources as insurance. These findings shed light on the design of incentive-compatible safety nets and development interventions. The paper is organised as follows. Section 2 describes our sampling strategy, our flood exposure variables, and the survey and summary statistics of our sampled households and villages. Section 3 discusses the empirical strategy we employed to identify causal impacts of the 2011 mega flood. Section 4 reports our empirical results. Section 5 concludes the paper with policy implications. 2. Data The data used in this study are from our survey conducted in April 2014 in four of Cambodia s key rice-growing provinces: Prey Veng, Kampong Thom, Banteay Meanchey and Battambang. As shown in Figure 4.1, these four provinces were severely affected by the 2011 flood. The four provinces also represent variations in geographical settings, rice cultivation and agricultural production systems, access to market opportunities, and the extent to which household livelihoods are prone to floods. These variations could potentially contribute to the variations in the nature of the 2011 flood experience, as well as the capacity and strategies of households and communities in coping with and managing floods. 88

6 Figure 4.1: Map of studied villages Inundated (>15 days) flood 2011 Sampled rice villages 2.1. Sampling strategy The survey and experiments cover 256 rice-farming households in 32 ricegrowing villages in 16 communes in the four provinces. Four considerations underlie our sampling strategy: First, we confine our study to rice growing areas and households. Second, we utilise the discontinuity generated by the 2011 flood to construct a variation in flood experience. This discontinuity allows us to compare villages and farmers directly hit by the flood with those who did not directly experience the flood. Third, spillover and generalequilibrium effects on the non-flood households were unavoidable. These effects include, but are not limited to, new information about the flood and the management of the flood by the government as perceived by the farmers. There were also disruptions to local, regional, and national economic activities that affected prices of goods and services, as well as incomes of many households in the non-flood areas. With household-level flood experience, the effects, however, should bias our results toward finding no difference in preferences 89

7 and behaviours between the farmers who were directly hit by the flood and those whose farms were not flooded. We also attempted to produce another set of comparable results to capture within-village spillover effects by creating variations in village-level 2011 flood experience. 3 Finally, since households in the flood-prone areas could have higher chance of being affected by the 2011 mega flood, relative to those in the non-flood-prone areas and the two groups could also have different characteristics, which could potentially result in different behavioural outcomes, simply selecting and comparing outcome variables between the flood affected and unaffected villages or households could leave us with a risk of selection problem leading our estimates to capture impacts of the flood risk rather than of the mega 2011 flood itself. Our sampling strategy, therefore, also involves further stratification by the degree to which households or villages are prone to floods to account for variations in flood risk, so that we can control for this problem outright in our econometric estimations. Overall, our sampling strategy for each province involves two stratifications, at both the village and household levels: (i) whether the village/household was flooded in 2011, and (ii) whether the village/household is generally prone to floods in normal years. To implement our sampling strategy, we went through the following steps. First, we used official statistics of rice production by commune and village from the Cambodian Council of Agricultural and Rural Development to identify our sampling frames in each province, i.e., the rice-producing communes and villages. We then used remote sensing maps of inundated areas produced by the World Food Program (WFP) to identify (i) communes severely affected by the 2011 mega flood (i.e., areas identified as inundated for more than 15 days) and (ii) communes that are prone to floods (based on 10 years of inundation data) in our four provinces. 4 For each province, we then selected 3 We note that our strategy thus will not capture the likely spillover effects within the flooded commune, district or even province. But with village-level flood experience, the communelevel spillovers should bias our results toward finding no effect. 4 The WFP flood maps were based on the near real time remote sensing NASA-MODIS product with 1-km resolution. The MODIS inundation maps have been available every 15 days since Mapping of severely affected areas was done by defining severely affected areas as those (non-permanent water) areas covered with floodwater for more than 15 days (i.e., where we saw water in at least two consecutive inundation maps). The WFP s flood risk mapping utilises 10 years of inundation flood maps and produces three flood priority classifications based on the 10-year flood frequency. The first, second and third priority flood zones consist of areas that experienced at least three, two or one extended flood(s) in 90

8 four rice-growing communes with extended areas severely affected by the 2011 flood, and two of which are flood-prone. In total, 16 flooded communes were selected, half of which are flood-prone. Within each commune, there could also be a variation in the flood experience across rice-growing villages, e.g., with respect to the share of areas/households affected. In the second step, we exploited this potential variation by defining flooded villages as villages with a majority of areas severely flooded (i.e., with large areas identified as inundated for more than 15 days). Using GIS village locators and the flood maps, we then selected two rice-growing villages one severely flooded and another not (severely) flooded in each commune. 5 Chiefs of the chosen communes were consulted to confirm our GIS-based classification and accessibility of the chosen villages. In cases where our chosen villages did not fit our categorisation, 6 we relied on commune chiefs and commune-level data for village selection instead. In particular, a ricegrowing village is classified as a flooded village if more than 50 percent of households reported rice production loss following the 2011 flood. In total, 32 rice-growing villages were selected. In sum, the sampling strategy up to this point thus allowed us to ensure the variation in village-level 2011 flood experience (severely flooded versus not [severely] flooded), as well as the variation of flood risk (flood-prone versus not flood-prone) within the flooded and non-flooded village groups. Within each village, there could also be sources of exogenous variations of the 2011 flood experience across households. Since our sampled households were rice farmers, the variation in the 2011 flood experience could relate closely to the extent that the flood affected rice production the variation of which then depended largely on the (largely exogenous) correlations between rice production cycle, timing of the flood, and flood severity (flood height and the the past ten years. We selected our flood-prone communes from the group of communes in the WFP s first flood priority. 5 Since the 2011 mega flood was largely covariate, it was not possible to find a completely non-flooded village. Our distinction of the flooded and non-flooded villages is thus the intensity of the 2011 flood extent, observed through share of areas/households affected by flood. Our village level flood impact analysis thus explores marginal variations in the village flood experience. 6 One of the key reasons is that the resolution of our flood maps could only allow accurate flood identification at commune level. 91

9 inundation period). 7 In the third step, we again exploited these potential variations by proceeding to generate variations in the 2011 flood experience at the household level. A household was classified as a flooded household if it reported that its rice fields were submerged by floodwater for longer than 15 days in In consultation with the village chiefs during subsequent field visits, we finally selected eight rice-growing households in each village applying the following criteria: (i) both flooded (rice fields were flooded) and non-flooded (rice fields were not flooded) households were selected for each village and (ii) the rice fields of the chosen households were geographically dispersed and varied in terms of the size of farm land. The sample size by province is shown in Panel A of Table 4.1. Note that, although we had originally intended to collect a balanced sample for flooded and non-flooded households, the sample size was largely unbalanced. The flooded households largely outnumbered non-flooded households for Kampong Thom, Banteay Meanchey and Battambang, where the majority of rice farms were flooded in Our samples were relatively more balanced in Prey Veng (29 flooded households out of 64 households). 7 It is possible that some of these factors could be correlated with household characteristics. For example, some advanced households may study and adjust their rice growing patterns to escape common floods. However, we argued that the majority of these factors were largely exogenous for Cambodian rice farmers. First, a large variation in the rice growing cycle was driven by variation in rice varieties. For example, long-life vs. short-life rice, or flooded vs. non-flooded rice are all common varieties in our studied areas. Second, while some farmers could learn to adjust their growing patterns to be more resilient to climate change, the extent and severity of the 2011 mega flood had been largely unexpected by rice farmers, as discussed in Section 1. In the survey, we also asked farmers if they had done anything to prepare for the 2011 flood; most answered that they had done nothing to prepare. 8 Using this definition, our estimation results using household-flood experience should capture flood impacts on households that had seen their rice production hit directly by the 2011 flood. A common occurrence were households that did not experience rice production damage even though housing and (bare) agricultural land were flooded, e.g., if they had harvested their rice prior to the flood. Such households we classified as non-flooded households. 92

10 Table 4.1: Sampling and Summary Statistics of the 2011 Mega Flood by Studied Province A. Sampled households Total villages Flooded villages Total households Flooded All Prey Veng Kampong Thom Banteay Meanchey Battambang All Prey Veng Kampong Thom Banteay Meanchey B. Characteristics of flood 2011 Mean SD Mean SD Mean SD Mean SD Mean SD Starting month Flood height Flood days Affected rice farm (%) Rice income lost (%) Consumption lost (%) Rice income lost ($) 1,648 6,150 1,459 1,693 1,209 4, ,425 11,050 Asset lost ($) 163 1, ,063 With house damage (%) With productive asset lost (%) With member lost (%) With reduced consumption (%) With reduced schooling (%) With reduced health care (%) C. Coping strategies Mean SD Mean SD Mean SD Mean SD Mean SD Forest clearance Collect forest product/fishing Asset sale Drawing out saving Child labor Adult labor Borrowing from banks Borrowing from MFIs, groups Borrowing from friends/relatives Borrowing amount ($) ,187 1, ,027 Remittances Governments NGOs Battambang All Prey Veng Kampong Thom Banteay Meanchey Battambang Flood height = 1 if very little, = 2 if knee high = 3 if chest high = 4 if above chest high. Coping strategies reported as percent of flooded households using the strategies. Figure 4.1 shows our survey villages in the four provinces overlaid with the 2011 flood map. Prey Veng is located in the southeastern plain on the crossing of the Upper Mekong and Lower Mekong rivers, the two major rivers in Cambodia. With annual flow of water from both rivers, the province is one of the high-potential agricultural zones of the country. Apart from rice, farmers 93

11 often diversify into other high-potential cash crops. The province also has good access to market and financial services due to its close proximity to the capital city, Phnom Penh. The other three provinces are located in the Tonle Sap Biosphere Reserve, meaning people there also greatly rely on the forest and natural resources for their livelihoods. Kampong Thom is located on the eastern floodplain of Tonle Sap lake and occupies key core biodiversity areas in the reserve. The province is among the largest in the country, so people have good access to employment and financial services. Banteay Meanchey occupies the extended lowland floodplain of Tonle Sap lake in the northwest. The province also has a border with Thailand and its people benefit from cross-border labour migration opportunities. Battambang is the country s largest rice production province in Cambodia and its rice is predominantly a high-yielding variety. The province also serves as a commercial and tourist hub in the northwestern region, with extended market access and alternative livelihoods, making the province wealthier than the other three. The 2011 mega flood posed a serious challenge to development and the livelihoods of people in all these four rice-growing provinces. The variations of flood experience across the four provinces are shown in Panel B of Table 4.1. Since the 2011 flood had resulted from the overflow of rainwater from the Mekong River toward Tonle Sap lake, it hit Prey Veng slightly earlier, in late August, before continuing to Kampong Thom, Banteay Meanchey, and Battambang in early September. Flood heights were also different with the majority of households in Prey Veng experiencing knee-high flood, whereas the other three provinces in the Tonle Sap region experienced chest-high flood. Households also reported the number of days that their rice fields were completely submerged by floodwater. We used this information to generate the total number of days that each household experienced the 2011 flood. 9 On average, the mega flood resulted in 26 submerged days, with a maximum of 180 days experienced in Kampong Thom. The mega flood damaged 89 percent of rice fields and resulted in an average of USD 1,648 lost in rice income and USD163 lost in assets in the four provinces, per household. The largest loss 9 We note that rice fields are typically located in lower land rather than in residential areas. If the housing areas were also flooded, it is very likely that the rice fields were also and still flooded. Thus, our household flood days could potentially capture the (non-linear) intensity of the 2011 flood, especially when the flood levels were high enough to damage housing and household assets. 94

12 was suffered by the relatively wealthy rice farmers in Battambang (averaging USD3,425 rice income loss and USD408 asset loss). Among the key assets lost were livestock and productive farm assets. Only seven percent of households reported damaged housing and one percent reported having lost family members. Following the 2011 flood, 24 percent of our sampled households reported they had to reduce consumption, nine percent had to cut back on child schooling, and 15 percent on health care, with slightly greater impacts in Kampong Thom. Panel C of Table 4.1 shows the variations of coping strategies the flooded households used during the 2011 mega flood across the four provinces. Strikingly, despite great variations, reliance on natural resources as a safety net was the most salient mechanism in all of the provinces it was adopted by 39 percent of flooded households. Social mechanisms and reliance on assistance from the government or non-governmental organisations (NGOs) were quite limited and varied greatly across the four provinces. Specifically, 22 percent of flooded households relied on remittances and borrowing from friends and relatives, although shares varied from only three percent in Prey Veng to 31 percent in Kampong Thom. Fifteen percent of flooded households relied on the government and 19 percent on NGOs, but the bulk of such assistance was concentrated in Kampong Thom. Apart from natural resources, our sampled rice-farming households relied more on various self-coping mechanisms 29 percent of flooded households reported using borrowing to cope with the 2011 flood, more than half of which borrowed from informal institutions such as microfinance institutions and saving groups. Use to credit to cope with the flood also varied across provinces, ranging from 45 percent in Prey Veng, 37 percent in Kampong Thom, 20 percent in Battambang, to 16 percent in Banteay Meanchey. Savings were used by some 24 percent of affected households and 27 percent of flooded households, especially in the three provinces in the Tonle Sap region, used additional labour income to cope with the 2011 flood. Despite the variety of strategies available, the use of destructive strategies, e.g., asset sales and child labour, were also common in some provinces. Overall, the above statistics suggest (i) significant and varying impacts of the 2011 flood on rice-farming communities in Cambodia; (ii) the importance of natural resources as a safety net during the mega flood; (iii) a striking limit to 95

13 social and government/ngos assistance during the flood; and (iv) the great extent and variety of self-coping mechanisms used by flooded Cambodian farmers during the flood. These varying flood experiences, opportunities and limits to the use of various mechanisms among affected households, therefore, could affect preferences, subjective expectations and behavioural choices The 2011 flood exposures Our sampling strategy discussed above allows us to construct three flood exposure variables. First, village-level flood exposure is a binary variable indicating whether the household was in a (relatively more severely) flooded village in 2011, where flooded village is defined as a village with a majority of areas flooded for more than 15 days and/or a village with more than 50 percent of households reporting rice production loss due to the flood. Employing this flood variable, our estimations should be able to identify the potential (marginal) impacts on households living in severely flooded villages relative to those living in not so severely flooded villages. Thus, the estimated impacts should generally include overall effects including likely spillover and general equilibrium effects on non-flooded households in these severely flooded villages. We note that our estimates could still suffer from the likely spillover effects within the flooded commune, district, province, or even country. But with village-level flood exposure, spillover effects at the higher levels should bias our results toward finding no effect. Second, household-level flood exposure is another binary variable indicating whether a household was flooded in 2011 (i.e., when their rice fields were completely submerged by floodwater for more than 15 days). Employing this household-level flood variable, our estimations should be able to identify the potential impacts on households directly hit by the 2011 flood. However, estimated impacts could still suffer from likely spillover effects, which again should bias out results toward finding no effect. Finally, we also used the number of days that households rice fields were completely submerged by floodwater to capture continuous household-level flood intensity. Our estimations using this flood variable should identify the potential heterogeneous effect of different levels of flood intensity on flooded 96

14 households. Altogether, these three variables should capture the varying aspects of the 2011 flood experienced by Cambodian rice-farming households The Survey The fieldwork conducted in April 2014 includes a standard household socioeconomic survey with detailed questions on the 2011 flood experience, other risks experienced by households over the past 10 years, risk management strategies, as well as key behavioural choices related to farm investment, saving and other safety net behaviours. The fieldwork also included a series of hypothetical experiment questions to elicit risk, time, social preferences; subjective expectations of future floods and resulting income loss; and household perceptions of the reliability of various safety net institutions to protect against the impacts of future floods. Appendix 1 provides a summary of the experiments and the associated preference parameters. First, for risk preference, we replicated the simple Binswanger (1980) game by allowing respondents to choose different rice seed types with different degrees of risk and return. Respondents seed choices could thus reflect their degree of risk aversion. We then constructed our risk aversion variable as a scaling indicator ranging from 1 (least averse) to 5 (most averse). Second, for time preference, the experiment consisted of a series of seven questions, each asking a respondent to choose between the choice of receiving some amount of money now or receiving a larger amount (that kept increasing as the experiment progressed from questions 1 to 7) in the future if he or she could wait to receive it. Observing the patterns of answers to these seven questions specifically the first time when the respondent chose to accept the payment in the future could reflect the extent to which respondents discount the future over the present, i.e., the degree of impatience. We then construct our impatience variable as a scaling indicator ranging from 0 (not impatient) to 8 (most impatient) We note that our simple measure of time preference is subject to risk aversion, as preferring to accept lower instantaneous payment to higher future payment may reflect an aversion to future payment that could be perceived as risky, in addition to time impatience. 97

15 Third, for social preference, we used a dictator game to illicit measures of household s altruism. Each respondent was given some amount of money, all or part of which they could give to a randomly chosen household in their village. The respondent was also told that the chosen beneficiary would be anonymous and that the respondent s decision would be kept confidential. We repeated this game but changed the beneficiary to be a randomly chosen floodaffected household in their village. We then constructed our altruism variable for each game from the proportion (0-100 percent) of money respondent chose to give. Fourth, in our experiments on subjective expectations we asked each respondent to assign probabilities to future flood events. We used 10 coins as visual aids to express the probability concept 11 and asked each respondent to place the coins in front of each of three flood events (no flood, mild flood, and mega flood), where the number of the coins he/she put would reflect the likelihood he/she thought each event would occur in the next 10 years. Before we began the exercise, our enumerator first clarified the definition of mild flood i.e., a flood event with less than knee-high floodwater and fewer than 10 days of waterlogging in the farm and the definition of severe flood i.e., a flood event with more than knee-high floodwater or more than 10 days of waterlogging in the farm and explained the exercise, using several examples (see Appendix 1). We repeated this exercise to also elicit the respondents perceptions of the likely proportion of rice income loss and the reliability of various safety nets conditional on the occurrence of mild and severe floods in the future. We then constructed each respondent s subjective expectation variables directly from the number of coins he/she assigned to each event. Finally, we also used a general social science survey to elicit the degrees to which each respondent trusted family, neighbours, businesses and local governments. These questions allowed us to construct series of binary trust variables. 11 Visual aids such as ours have been used widely in low-income countries with relatively illiterate subjects who may find direct questions about probability too abstract. See Delavande, et al. (2011) for a review. 98

16 2.4. Summary statistics of sampled households Table 4.2 reports descriptive statistics of the sampled households by village and household-level flood exposure at the time of the survey in April Overall, household and village characteristics were similar for flooded versus non-flooded villages, and especially for flooded versus non-flooded households. The average household size was about five people. Seventy-eight percent of respondents in the flooded households had primary education, 32 percent had secondary education and these statistics were not significantly different for non-flooded households. Average land owned was 0.53 hectare for flooded households with a mean income per capita of USD per year, 47 percent of which came from rice production. About 23 percent of flooded households were classified as poor according to the Identification of Poor Household Program (ID Poor) and had faced about 2.3 other shocks over the past 10 years. Again, these statistics were similar for the non-flooded group. Availability of key village infrastructure and public programmes also appeared similar across flood groups. Table 4.2 also shows some characteristics that were significantly different between the flooded and non-flooded villages e.g., gender of the respondents, household size and land per capita. We constructed a flood-prone variable from the frequency of floods reported by each household and so a household was prone to floods if it reported at least two floods experiences in the past five years. Our statistics also shows that flooded households were significantly more flood-prone than non-flooded households, with an average flood frequency of 1.75 in the past five years. If the key characteristics we found to be different across flood groups were also correlated with our behavioural outcomes of interest, this could potentially bias our estimation results. It is important, therefore, that we control for these variables in our empirical analysis. 99

17 Table 4.2: Summary Statistics of Sampled Households by Flood Exposure Household characteristics Village flood (=1) Household flood (=1) Flooded Not flooded Difference Flooded Not flooded Difference Female (=1) *** (0.477) (0.502) (0.061) (0.497) (0.488) (0.065) Age (12.29) (13.04) (1.583) (12.70) (12.57) (1.685) Have education-primary (=1) ** (0.365) (0.443) (0.051) (0.416) (0.395) (0.054) Have education-secondary (=1) (0.482) (0.459) (0.059) (0.467) (0.478) (0.062) Household size ** (2.238) (1.652) (0.245) (2.070) (1.777) (0.263) Member migrate (%) (1.159) (0.945) (0.132) (1.069) (1.033) (0.140) Female member migrate (%) (0.656) (0.485) (0.072) (0.605) (0.516) (0.076) Age of migrating members (27.97) (25.52) (3.347) (27.24) (25.67) (3.560) Income per capita ($) (903.81) ( ) (198.88) ( ) (706.81) (211.67) Rice income in total income (%) (0.349) (0.357) (0.044) (0.345) (0.361) (0.046) Land per capita (ha) * (0.774) (0.506) (0.081) (0.684) (0.595) (0.087) Asset per capita ($) ( ) ( ) (384.23) ( ) ( ) (410.78) ID poor household (=1) (0.415) (0.434) (0.053) (0.423) (0.428) (0.056) Flood prone (=1) *** (0.500) (0.491) (0.061) (0.484) (0.500) (0.065) Flood frequency in the past 5 yrs *** (0.774) (0.763) (0.096) (0.612) (0.915) (0.096) Other shocks in the past 10 yrs (1.674) (1.829) (0.219) (1.651) (1.932) (0.232) Village characteristics Have irrigation system (=1) (0.516) (0.466) (0.057) (0.459) (0.470) (0.061) Have electricity (=1) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) With social land concession (=1) (0.313) (0.281) (0.037) (0.314) (0.259) (0.039) With health equity fund (=1) (0.409) (0.322) (0.046) (0.394) (0.311) (0.049) Standard deviations in parentheses. * p<0.1; ** p<0.05; *** p<

18 Table 4.3 reports descriptive statistics of our measures of preferences, subjective expectations, and behavioural choices, again, at the time of the survey in April The table shows that the sampled households were relatively risk averse with both the mean and the median measures of risk aversion ranging from in all groups. Our simple comparison showed that the mean risk aversion variables were not significantly different between flooded and non-flooded villages or households. Figure 4.2 plots distributions of the risk aversion parameter by household flood experience. These plots provide the additional finding that the share of households with extreme risk aversion appeared larger among the flooded households. 101

19 Table 4.3: Summary Statistics of Preference and Behavioral Variables by Flood Exposure Village flood (=1) Household flood (=1) Flooded Not flooded Difference Flooded Not flooded Difference Preferences Risk aversion (1,2,..,5) (1.473) (1.425) (0.181) (1.474) (1.389) (0.192) Impatience (0,1,2,..,8) * (2.635) (2.593) (0.326) (2.643) (2.511) (0.346) Altruism - percent money given to randomly matched ** ** villager (0-1) (0.239) (0.202) (0.027) (0.234) (0.192) (0.029) Altruism - percent money given to randomly matched ** * flood victim in the village (0-1) (0.245) (0.192) (0.027) (0.232) (0.198) (0.030) Trust family (=1) (0.088) (0.124) (0.013) (0.131) (0.000) (0.014) Trust neighbor (=1) ** (0.332) (0.340) (0.042) (0.385) (0.153) (0.043) Trust business/trader (=1) * (0.496) (0.476) (0.060) (0.487) (0.491) (0.065) Trust local government (=1) ** (0.420) (0.439) (0.053) (0.449) (0.374) (0.056) Subjective expectations Probability of mild flood (0-1) (0.228) (0.225) (0.028) (0.224) (0.230) (0.030) Probability of severe flood (0-1) *** (0.262) (0.262) (0.032) (0.263) (0.244) (0.034 Probability of loss when mild flood occurs (0-1) *** (0.282) (0.285) (0.035) (0.291) (0.243) (0.036) Probability of loss when severe flood occurs (0-1) (0.286) (0.260) (0.034) (0.222) (0.342) (0.035) Can rely on govnt. when mild flood (=1) (0.232) (0.220) (0.028) (0.226) (0.226) (0.030) Can rely on govnt. when severe flood (=1) (0.310) (0.300) (0.038) (0.295) (0.326) (0.040) Can rely on social network when mild flood (=1) * *** (0.260) (0.272) (0.033) (0.297) (0.176) (0.035) Can rely on social network when severe flood (=1) * * (0.237) (0.280) (0.032) (0.275) (0.222) (0.034) Can rely on natural resource when mild flood (=1) (0.372) (0.350) (0.045) (0.351) (0.382) (0.048) Can rely on natural resource when severe flood (=1) (0.345) (0.306) (0.040) (0.328) (0.322) (0.043) Behavioral choices Investment in land and irrigation (=1) (0.349) (0.322) (0.042) (0.328) (0.352) (0.044) Have saving (=1) *** (0.392) (0.392) (0.049) (0.433) (0.259) (0.051) Number of dependable friends (1.049) (0.822) (0.118) (0.933) (0.965) (0.126) Collect forest products and fishing (=1) ** (0.281) (0.313) (0.037) (0.265) (0.352) (0.394) Demand market insurance (=1) ** Standard deviations in parentheses. * p<0.1; ** p<0.05; *** p<0.01 (0.293) (0.281) (0.035) (0.314) (0.214) (0.038) 102

20 Figure 4.2: Risk Aversion, Impatience, Altruism and Trust by Household Flood Exposure The impatience variable appeared similar between households in flooded versus non-flooded villages. Our simple comparison, however, shows that flooded households seemed to be significantly less impatient than non-flooded households. Figure 4.2 further shows that the share of households with extreme impatience appeared smaller among flooded households than among the nonflooded group. 103

21 On average, there appeared to be significantly larger altruism variables for flooded households and households in flooded villages than for non-flooded groups. The average share of money given to a randomly matched villager was about 0.25 in the flooded group. As shown in Figure 4.2, a smaller share of households gave nothing but a larger share of households gave a large amount to a random villager in the flooded group than that of the non-flooded group. And in all groups, the proportion given to a random villager was smaller than that given to a flood victim. For trust, we found that in all groups almost all (99 percent) of our sampled households trusted family, followed by trusting neighbours (82-98 percent), trusting local governments (72-83 percent) and trusting businesses (34-43 percent). The share of households that trusts family and businesses appears similar across flood groups, whereas the share of those trusting neighbours and local government appears significantly smaller in the flooded group. For subjective expectations, our sampled households assigned large probabilities of flood risk in general ( for mild flood and for severe flood). This was to be expected given that our samples are all from flood-affected communes. The flooded households, however, assigned significantly higher subjective probabilities to severe flood, and also a significantly higher perceived proportion of rice income loss in the event of a mild flood. Finally, the descriptive statistics of households perceptions on safety net institutions also revealed some interesting results among our sampled ricefarming households. For both mild and severe floods, the largest percentage of households (27-37 percent) in all groups perceived that they could rely on natural resources as a safety net. These were followed by a perceived ability to rely on governments (28-30 percent) and social networks (12-17 percent) when a severe flood occurs. For mild flood, however, both perceived ability to rely on governments and social networks appeared to be similar, at only percent. Statistically, these safety net perceptions were not significantly different across flood groups, except for the perceived ability to rely on social networks. Similar findings are depicted in Figure

22 Figure 4.3: Subjective Expectations by Household Flood Exposure We are also interested in the potential impacts of the 2011 mega flood on some key behavioural choices that could potentially determine households economic growth and their resilience to future floods. The variables of our interest are (i) whether a household invested in land and irrigation; (ii) whether a household had savings; (iii) the number of dependable friends a household had (as an indicator of social capital formation); (iv) whether a household collected forest products and engaged in fishing; and (v) a household s willingness to pay for commercial flood insurance. Interestingly, Table

23 reveals that a significantly larger percentage of households had savings and demand for commercial insurance, and a significantly smaller percentage of households collected forest products among flooded households than among non-flooded households. These bivariate relationships in Table 4.3, however, should be interpreted with some caution. To what extent might these relationships be driven by other observed and/or unobserved variables that were correlated with both 2011 flood exposure and our outcome variables? Figure 4.4 depicts some bivariate relationships between our preference and expectation variables and (i) whether a household was flood-prone; (ii) land ownership; and (iii) education the key covariate theoretically known to affect these behavioural variables. As expected, these figures suggest that risk aversion was positively associated with the degree of flood risk and negatively associated with wealth and education. Altruism also appeared to increase with flood risk and wealth. And the subjective probabilities of future floods were also positively associated with the degree of flood risk. Since some of these key variables were also correlated with flood exposure (e.g., flood-prone and land ownership), we will control for these variables in our estimations in the next section. 106

24 Figure 4.4: Relationships between Preferences and Key Characteristics 107

25 3. Empirical Strategy We estimate the potential impacts of the 2011 mega flood by regressing our preference and behavioural variables on flood exposure, controlling for individual, geographical characteristics, and village fixed effects. Our estimations thus follow a simple specification: y iv = β 0 + β 1 Flood iv + β 2 Flood iv Floodprone iv + β 3 X iv + α v + ε iv where y iv represents preference, subjective expectation, or other behavioural choice variables of interest. Flood iv is a variable that captures households exposure to the 2011 flood. In our analysis, we use three different measures of this flood exposure: (i) a village-level indicator if a household was in the flooded village, utilising the exogenous variation of flood experience across villages; (ii) a household-level indicator if a household was directly affected by flood, utilising exogenous variations of flood experience across households within each village; and (iii) the number of days that a household s rice fields were completely submerged by floodwater, capturing the continuous household-level flood intensity. Floodprone iv is a household-level indicator variable controlling for the potential lurking effect of the degree to which each household was prone to floods. 12 X iv are various household-level controls while α v controls for unobserved heterogeneity at village level. 13 We also clustered all specifications at the commune level. Various potential sources of selection bias are worth discussing. First, one would wonder if the variations of village-level flood experience were exogenous. Since the flood-prone villages were likely be flooded, the floodprone variable would be correlated with some key behavioural variables. To address this concern, we stratify our sample by their vulnerability to flood, captured by the flood-prone variable, and control for this in the estimation. Another potential problem is migration, which could generate an endogeneity in flood exposure, especially if many households moved from flooded to nonflooded areas. However, this problem should be minimal for our sampled 12 Again, flood-prone equals one if household had experienced at least two floods over the past five years. 13 For village flood exposure, commune level fixed effect was used. 108